nsz Git - libfirm/blob - ir/be/bespillbelady3.c

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       MIN-algorithm with some twists (aka belady spiller v3)
  23  * @author      Matthias Braun
  24  * @date        15.01.2008
  25  * @version     $Id$
  26  *
  27  * TODO:
  28  *   - handle phis correctly, decide whether we should spill them ~ok
  29  *   - merge multiple start worksets of blocks ~ok
  30  *   - how to and when to do the tentative phase...
  31  */
  32 #include "config.h"
  33
  34 #include <stdbool.h>
  35
  36 #include "debug.h"
  37 #include "list.h"
  38 #include "pdeq.h"
  39
  40 #include "irnode_t.h"
  41 #include "irprintf.h"
  42 #include "iredges_t.h"
  43 #include "irloop_t.h"
  44 #include "irgwalk.h"
  45 #include "execfreq.h"
  46
  47 #include "bemodule.h"
  48 #include "bespill.h"
  49 #include "beutil.h"
  50 #include "bespillutil.h"
  51 #include "besched.h"
  52 #include "be_t.h"
  53
  54 #ifndef NDEBUG
  55 #define EXPENSIVE_CHECKS
  56 #endif
  57
  58 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  59
  60 typedef struct loop_edge_t loop_edge_t;
  61 struct loop_edge_t {
  62         ir_node     *block;
  63         int          pos;
  64         loop_edge_t *next;
  65 };
  66
  67 typedef struct loop_info_t loop_info_t;
  68 struct loop_info_t {
  69         ir_loop     *loop;
  70         loop_edge_t *entry_edges;
  71         loop_edge_t *exit_edges;
  72         size_t       max_register_pressure;
  73 };
  74
  75 typedef struct worklist_entry_t worklist_entry_t;
  76 struct worklist_entry_t {
  77         struct list_head  head;
  78         ir_node          *value;
  79         ir_node          *reload_point;
  80         ir_loop          *unused_livethrough_loop;
  81 };
  82
  83 #define foreach_worklist(entry, wl) list_for_each_entry(worklist_entry_t, entry, &(wl)->live_values, head)
  84
  85 typedef struct worklist_t worklist_t;
  86 struct worklist_t {
  87         struct list_head  live_values;
  88         size_t            n_live_values;
  89         ir_visited_t      visited;
  90 };
  91
  92 typedef struct block_info_t block_info_t;
  93 struct block_info_t {
  94         worklist_t *start_worklist;
  95         worklist_t *end_worklist;
  96 };
  97
  98 /** The register class we currently run on. */
  99 static const arch_register_class_t *cls;
 100 static struct obstack               obst;
 101 static spill_env_t                 *senv;
 102 static size_t                       n_regs;
 103 static size_t                       max_register_pressure;
 104 static bool                         tentative_mode;
 105 static bool                         do_push_unused_livethroughs;
 106 /** Execution frequency for the current graph. */
 107 static ir_exec_freq                *exec_freq;
 108 static ir_visited_t                 worklist_visited;
 109 #ifdef DEBUG_libfirm
 110 static bool                         should_have_reached_fixpoint;
 111 #endif
 112
 113 static worklist_t *new_worklist(void)
 114 {
 115         worklist_t *worklist = OALLOCZ(&obst, worklist_t);
 116
 117         INIT_LIST_HEAD(&worklist->live_values);
 118         worklist->n_live_values    = 0;
 119
 120         return worklist;
 121 }
 122
 123 static void mark_irn_not_visited(ir_node *node)
 124 {
 125         set_irn_visited(node, get_irg_visited(current_ir_graph) - 1);
 126 }
 127
 128 static bool worklist_contains(const ir_node *node)
 129 {
 130         return irn_visited(node);
 131 }
 132
 133 static block_info_t *get_block_info(ir_node *block)
 134 {
 135         block_info_t *info = get_irn_link(block);
 136         if (info != NULL)
 137                 return info;
 138
 139         info = OALLOCZ(&obst, block_info_t);
 140         set_irn_link(block, info);
 141         return info;
 142 }
 143
 144 static void deactivate_worklist(const worklist_t *worklist)
 145 {
 146         worklist_entry_t *entry;
 147
 148         foreach_worklist(entry, worklist) {
 149                 assert(worklist_contains(entry->value));
 150                 mark_irn_not_visited(entry->value);
 151                 set_irn_link(entry->value, NULL);
 152         }
 153 }
 154
 155 static void activate_worklist(const worklist_t *worklist)
 156 {
 157         worklist_entry_t *entry;
 158
 159         foreach_worklist(entry, worklist) {
 160                 assert(!worklist_contains(entry->value));
 161                 mark_irn_visited(entry->value);
 162                 set_irn_link(entry->value, entry);
 163         }
 164 }
 165
 166 /**
 167  * duplicate a worklist and directly activates it
 168  */
 169 static void fill_and_activate_worklist(worklist_t *new_worklist,
 170                 const worklist_t *worklist, ir_node *block, ir_node *succ_block,
 171                 int succ_pos)
 172 {
 173         ir_node          *reload_point  = NULL;
 174         size_t            n_live_values = 0;
 175         worklist_entry_t *entry;
 176
 177         if (succ_block != NULL &&
 178                         (get_Block_n_cfgpreds(succ_block) > 1
 179                          || get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) > 1)) {
 180                 reload_point = be_get_end_of_block_insertion_point(block);
 181         }
 182
 183         foreach_worklist(entry, worklist) {
 184                 ir_node          *value = entry->value;
 185                 worklist_entry_t *new_entry;
 186
 187                 if (new_worklist->n_live_values >= n_regs)
 188                         break;
 189
 190                 if (is_Phi(value) && get_nodes_block(value) == succ_block) {
 191                         value = get_Phi_pred(value, succ_pos);
 192
 193                         /* can happen for unknown phi preds */
 194                         if (!arch_irn_consider_in_reg_alloc(cls, value))
 195                                 continue;
 196                 }
 197
 198                 if (irn_visited_else_mark(value))
 199                         continue;
 200
 201                 new_entry = OALLOCZ(&obst, worklist_entry_t);
 202
 203                 new_entry->value = value;
 204                 if (reload_point != NULL) {
 205                         new_entry->reload_point = reload_point;
 206                 } else {
 207                         new_entry->reload_point = entry->reload_point;
 208                 }
 209
 210                 list_add_tail(&new_entry->head, &new_worklist->live_values);
 211                 ++n_live_values;
 212
 213                 set_irn_link(value, new_entry);
 214                 new_worklist->n_live_values++;
 215         }
 216 }
 217
 218 static worklist_t *duplicate_worklist(const worklist_t *worklist)
 219 {
 220         worklist_t       *new_worklist;
 221         struct list_head *entry;
 222
 223         new_worklist = OALLOCZ(&obst, worklist_t);
 224         INIT_LIST_HEAD(&new_worklist->live_values);
 225         new_worklist->n_live_values = worklist->n_live_values;
 226
 227         list_for_each(entry, &worklist->live_values) {
 228                 worklist_entry_t *wl_entry  = list_entry(entry, worklist_entry_t, head);
 229                 worklist_entry_t *new_entry     = OALLOC(&obst, worklist_entry_t);
 230
 231                 memcpy(new_entry, wl_entry, sizeof(new_entry[0]));
 232                 list_add_tail(&new_entry->head, &new_worklist->live_values);
 233         }
 234
 235         return new_worklist;
 236 }
 237
 238 #ifdef DEBUG_libfirm
 239 static void print_worklist(const worklist_t *worklist, int level)
 240 {
 241         struct list_head *entry;
 242
 243         DB((dbg, level, "%d values: ", worklist->n_live_values));
 244         list_for_each(entry, &worklist->live_values) {
 245                 worklist_entry_t *wl_entry
 246                         = list_entry(entry, worklist_entry_t, head);
 247
 248                 DB((dbg, level, "%+F ", wl_entry->value));
 249         }
 250 }
 251 #endif
 252
 253
 254 /**
 255  * Clear the link fields of a loop and all
 256  * its child loops.
 257  *
 258  * @param loop  the root loop
 259  */
 260 static void clear_loop_info(ir_loop *loop)
 261 {
 262         int n_elements = get_loop_n_elements(loop);
 263         int i;
 264
 265         loop->link = NULL;
 266         for (i = 0; i < n_elements; ++i) {
 267                 loop_element element = get_loop_element(loop, i);
 268                 if (*element.kind != k_ir_loop)
 269                         continue;
 270
 271                 clear_loop_info(element.son);
 272         }
 273 }
 274
 275 static loop_info_t *get_loop_info(ir_loop *loop)
 276 {
 277         loop_info_t *info = loop->link;
 278         if (info != NULL)
 279                 return info;
 280
 281         info = OALLOCZ(&obst, loop_info_t);
 282         info->loop = loop;
 283         loop->link = info;
 284         return info;
 285 }
 286
 287 /**
 288  * constructs loop in+out edges when called in a block walker
 289  */
 290 static void construct_loop_edges(ir_node *block, void *data)
 291 {
 292         int      n_cfgpreds = get_Block_n_cfgpreds(block);
 293         ir_loop *loop       = get_irn_loop(block);
 294         int      i;
 295
 296         (void) data;
 297
 298         for (i = 0; i < n_cfgpreds; ++i) {
 299                 ir_node     *cfgpred_block = get_Block_cfgpred_block(block, i);
 300                 ir_loop     *cfgpred_loop  = get_irn_loop(cfgpred_block);
 301                 loop_edge_t *edge;
 302                 bool        is_exit_edge;
 303                 ir_loop     *l, *goal;
 304
 305                 if (cfgpred_loop == loop)
 306                         continue;
 307
 308                 /* critical edges are splitted, so we can't jump from 1 loop
 309                  * directly into another without going out of it */
 310                 assert(get_loop_depth(cfgpred_loop) != get_loop_depth(loop));
 311
 312                 /* edge out of loop */
 313                 if (get_loop_depth(cfgpred_loop) > get_loop_depth(loop)) {
 314                         is_exit_edge = true;
 315                         l            = cfgpred_loop;
 316                         goal         = loop;
 317                 } else {
 318                         is_exit_edge = false;
 319                         l            = loop;
 320                         goal         = cfgpred_loop;
 321                 }
 322
 323                 /* this might be a jump out/in multiple loops */
 324                 do {
 325                         loop_info_t *l_info = get_loop_info(l);
 326
 327                         edge = OALLOCZ(&obst, loop_edge_t);
 328                         edge->block = block;
 329                         edge->pos   = i;
 330
 331                         if (is_exit_edge) {
 332                                 edge->next         = l_info->exit_edges;
 333                                 l_info->exit_edges = edge;
 334                                 assert(get_loop_depth(loop) < get_loop_depth(l));
 335                         } else {
 336                                 edge->next          = l_info->entry_edges;
 337                                 l_info->entry_edges = edge;
 338                                 assert(get_loop_depth(cfgpred_loop) < get_loop_depth(l));
 339                         }
 340                         l = get_loop_outer_loop(l);
 341                 } while (l != goal);
 342         }
 343 }
 344
 345
 346
 347
 348 static void place_reload(worklist_entry_t *entry)
 349 {
 350         if (tentative_mode)
 351                 return;
 352
 353         DB((dbg, LEVEL_1, "reload of %+F before %+F\n", entry->value,
 354                 entry->reload_point));
 355         assert(entry->reload_point != NULL);
 356         be_add_reload(senv, entry->value, entry->reload_point, cls, 1);
 357         entry->reload_point = NULL;
 358 }
 359
 360 /**
 361  * makes sure the worklist contains not more than n_regs - room_needed entries
 362  */
 363 static void make_room(worklist_t *worklist, size_t room_needed)
 364 {
 365         int               i;
 366         int               spills_needed;
 367         struct list_head *entry;
 368
 369         spills_needed = worklist->n_live_values + room_needed - n_regs;
 370         if (spills_needed <= 0)
 371                 return;
 372
 373         entry = worklist->live_values.next;
 374         for(i = spills_needed; i > 0; --i) {
 375                 struct list_head *next = entry->next;
 376                 worklist_entry_t *wl_entry
 377                         = list_entry(entry, worklist_entry_t, head);
 378
 379                 assert(worklist_contains(wl_entry->value));
 380                 mark_irn_not_visited(wl_entry->value);
 381                 place_reload(wl_entry);
 382                 list_del(entry);
 383
 384                 entry = next;
 385         }
 386         worklist->n_live_values -= spills_needed;
 387 }
 388
 389 /**
 390  * a value was used, so bring it to the back of the worklist (which might
 391  * result in a spill of another value).
 392  */
 393 static void val_used(worklist_t *worklist, ir_node *value, ir_node *sched_point)
 394 {
 395         /* already in the worklist? move around, otherwise add at back */
 396         worklist_entry_t *entry = get_irn_link(value);
 397
 398         assert(arch_irn_consider_in_reg_alloc(cls, value));
 399
 400         if (worklist_contains(value)) {
 401                 assert(entry != NULL);
 402
 403                 list_del(&entry->head);
 404         } else {
 405                 if (entry == NULL) {
 406                         entry        = OALLOCZ(&obst, worklist_entry_t);
 407                         entry->value = value;
 408                         set_irn_link(value, entry);
 409                 }
 410
 411                 ++worklist->n_live_values;
 412                 mark_irn_visited(value);
 413         }
 414
 415         entry->reload_point = sched_point;
 416         list_add_tail(&entry->head, &worklist->live_values);
 417 }
 418
 419 static void worklist_remove(worklist_t *worklist, ir_node *value)
 420 {
 421         worklist_entry_t *entry = get_irn_link(value);
 422         assert(entry != NULL);
 423         list_del(&entry->head);
 424         --worklist->n_live_values;
 425
 426         assert(worklist_contains(value));
 427         mark_irn_not_visited(value);
 428 }
 429
 430 static void update_max_pressure(worklist_t *worklist)
 431 {
 432         if (worklist->n_live_values > max_register_pressure)
 433                 max_register_pressure = worklist->n_live_values;
 434 }
 435
 436 static void do_spilling(ir_node *block, worklist_t *worklist)
 437 {
 438         ir_node *node;
 439
 440         assert(worklist != NULL);
 441
 442         sched_foreach_reverse(block, node) {
 443                 int    i, arity;
 444                 size_t n_defs = 0;
 445
 446                 DB((dbg, LEVEL_2, "\t%+F... ", node));
 447                 update_max_pressure(worklist);
 448
 449                 if (is_Phi(node)) {
 450                         ir_node *node2;
 451                         /* TODO: if we have some free registers, then we could decide to
 452                          * not spill some phis (but not for phis where at least 1 input is
 453                          * themselfes) */
 454
 455                         /* we have to spill all phis that are not live */
 456                         sched_foreach_reverse_from(node, node2) {
 457                                 assert(is_Phi(node2));
 458
 459                                 if (worklist_contains(node2))
 460                                         continue;
 461                                 if (!arch_irn_consider_in_reg_alloc(cls, node2))
 462                                         continue;
 463
 464                                 if (!tentative_mode)
 465                                         be_spill_phi(senv, node2);
 466                         }
 467                         DB((dbg, LEVEL_2, "\n"));
 468                         break;
 469                 }
 470
 471                 /* remove values defined by this instruction from the workset. Values
 472                  * defined but not in the workset need free registers */
 473                 if (get_irn_mode(node) == mode_T) {
 474                         const ir_edge_t *edge;
 475
 476                         foreach_out_edge(node, edge) {
 477                                 ir_node *proj = get_edge_src_irn(edge);
 478                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 479                                         continue;
 480                                 if (worklist_contains(proj)) {
 481                                         worklist_remove(worklist, proj);
 482                                 } else {
 483                                         ++n_defs;
 484                                 }
 485                         }
 486                 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
 487                         if (worklist_contains(node)) {
 488                                 worklist_remove(worklist, node);
 489                         } else {
 490                                 n_defs = 1;
 491                         }
 492                 }
 493
 494                 /* make sure we have enough free registers for the spills */
 495                 make_room(worklist, n_defs);
 496
 497                 /* put all values used by the instruction into the workset */
 498                 arity = get_irn_arity(node);
 499                 for(i = 0; i < arity; ++i) {
 500                         ir_node *use = get_irn_n(node, i);
 501
 502                         if (!arch_irn_consider_in_reg_alloc(cls, use))
 503                                 continue;
 504
 505                         val_used(worklist, use, node);
 506                 }
 507
 508                 /* we might have too many values in the worklist now and need to spill
 509                  * some */
 510                 make_room(worklist, 0);
 511
 512 #ifdef DEBUG_libfirm
 513                 print_worklist(worklist, LEVEL_2);
 514                 DB((dbg, LEVEL_2, "\n"));
 515 #endif
 516         }
 517
 518         update_max_pressure(worklist);
 519
 520 #ifdef DEBUG_libfirm
 521         DB((dbg, LEVEL_1, "worklist at begin of %+F:", block));
 522         print_worklist(worklist, LEVEL_1);
 523         DB((dbg, LEVEL_1, "\n"));
 524 #endif
 525 }
 526
 527 static bool worklists_equal(const worklist_t *wl1, const worklist_t *wl2)
 528 {
 529         const struct list_head *i1 = &wl1->live_values;
 530         const struct list_head *i2 = &wl2->live_values;
 531
 532         for ( ; i1 != &wl1->live_values && i2 != &wl2->live_values;
 533                         i1 = i1->next, i2 = i2->next) {
 534                 worklist_entry_t *entry1 = list_entry(i1, worklist_entry_t, head);
 535                 worklist_entry_t *entry2 = list_entry(i2, worklist_entry_t, head);
 536
 537                 if (entry1->value != entry2->value)
 538                         return false;
 539         }
 540         /* both lists at the end */
 541         if (i1 != &wl1->live_values || i2 != &wl2->live_values)
 542                 return false;
 543
 544         return true;
 545 }
 546
 547 static bool fill_start_worklist(worklist_t *new_worklist, ir_node *block)
 548 {
 549         double           best_execfreq   = -1;
 550         worklist_t      *best_worklist   = NULL;
 551         ir_node         *best_succ_block = NULL;
 552         int              best_pos        = -1;
 553         const ir_edge_t *edge;
 554
 555         /* construct worklist */
 556         foreach_block_succ(block, edge) {
 557                 ir_node      *succ_block = get_edge_src_irn(edge);
 558                 double       execfreq    = get_block_execfreq(exec_freq, succ_block);
 559                 block_info_t *block_info;
 560                 worklist_t   *succ_worklist;
 561
 562                 if (execfreq < best_execfreq)
 563                         continue;
 564
 565                 block_info    = get_block_info(succ_block);
 566                 succ_worklist = block_info->start_worklist;
 567
 568                 if (succ_worklist == NULL || succ_worklist->visited >= worklist_visited)
 569                         continue;
 570
 571                 best_execfreq   = execfreq;
 572                 best_worklist   = succ_worklist;
 573                 best_succ_block = succ_block;
 574                 best_pos        = get_edge_src_pos(edge);
 575         }
 576
 577         if (best_worklist == NULL)
 578                 return false;
 579         best_worklist->visited = worklist_visited;
 580
 581         fill_and_activate_worklist(new_worklist, best_worklist, block,
 582                         best_succ_block, best_pos);
 583         return true;
 584 }
 585
 586 static worklist_t *construct_start_worklist(ir_node *block)
 587 {
 588         worklist_t *worklist = new_worklist();
 589
 590         ++worklist_visited;
 591
 592         while(fill_start_worklist(worklist, block)) {
 593                 if (worklist->n_live_values >= n_regs)
 594                         break;
 595         }
 596
 597         return worklist;
 598 }
 599
 600 static void process_block(ir_node *block, void *env)
 601 {
 602         block_info_t    *block_info;
 603         worklist_t      *worklist;
 604         int              n_preds;
 605
 606         (void) env;
 607         DB((dbg, LEVEL_1, "Processing %+F\n", block));
 608
 609         worklist   = construct_start_worklist(block);
 610         block_info = get_block_info(block);
 611
 612 #ifdef DEBUG_libfirm
 613         DB((dbg, LEVEL_1, "worklist at end of %+F:", block));
 614         print_worklist(worklist, LEVEL_1);
 615         DB((dbg, LEVEL_1, "\n"));
 616
 617         if (should_have_reached_fixpoint) {
 618                 assert(worklists_equal(block_info->end_worklist, worklist));
 619         }
 620 #endif
 621         block_info->end_worklist = duplicate_worklist(worklist);
 622
 623         do_spilling(block, worklist);
 624         deactivate_worklist(worklist);
 625
 626 #ifdef DEBUG_libfirm
 627         if (should_have_reached_fixpoint) {
 628                 assert(worklists_equal(block_info->start_worklist, worklist));
 629         }
 630 #endif
 631         block_info->start_worklist = worklist;
 632
 633         /* we shouldn't have any live values left at the start block */
 634         n_preds = get_Block_n_cfgpreds(block);
 635         //assert(n_preds != 0 || worklist->n_live_values == 0);
 636 }
 637
 638 typedef union block_or_loop_t block_or_loop_t;
 639 union block_or_loop_t {
 640         ir_node *block;
 641         ir_loop *loop;
 642 };
 643
 644 static block_or_loop_t *loop_blocks;
 645 static ir_loop         *current_loop;
 646
 647 static void find_blocks(ir_node *block);
 648
 649 static void find_in_loop(ir_loop *loop, ir_node *entry)
 650 {
 651         loop_info_t     *loop_info = get_loop_info(loop);
 652         block_or_loop_t block_or_loop;
 653         loop_edge_t     *edge;
 654
 655         /* simply mark 1 block in the loop to indicate that the loop was already
 656          * processed */
 657         ir_node *some_block = loop_info->entry_edges->block;
 658         if (Block_block_visited(some_block))
 659                 return;
 660
 661         block_or_loop.loop = loop;
 662         ARR_APP1(block_or_loop_t, loop_blocks, block_or_loop);
 663
 664 #ifndef NDEBUG
 665         {
 666                 /* we should be 1 of the entry blocks */
 667                 loop_edge_t *edge  = loop_info->entry_edges;
 668                 bool         found = false;
 669                 for ( ; edge != NULL; edge = edge->next) {
 670                         if (edge->block == entry) {
 671                                 found = true;
 672                                 break;
 673                         }
 674                 }
 675                 assert(found);
 676         }
 677 #else
 678         (void) entry;
 679 #endif
 680
 681         /* check all loop successors */
 682         for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
 683                 ir_node *succ      = edge->block;
 684                 ir_loop *succ_loop = get_irn_loop(succ);
 685
 686                 if (succ_loop == current_loop) {
 687                         find_blocks(succ);
 688                 } else {
 689                         assert(get_loop_depth(succ_loop) < get_loop_depth(current_loop));
 690                 }
 691         }
 692 }
 693
 694 static void find_blocks(ir_node *block)
 695 {
 696         const ir_edge_t *edge;
 697         block_or_loop_t block_or_loop;
 698
 699         if (Block_block_visited(block))
 700                 return;
 701
 702         block_or_loop.block = block;
 703
 704         ARR_APP1(block_or_loop_t, loop_blocks, block_or_loop);
 705         mark_Block_block_visited(block);
 706
 707         foreach_block_succ(block, edge) {
 708                 ir_node *succ = get_edge_src_irn(edge);
 709
 710                 /* is it part of the current loop? */
 711                 ir_loop *loop = get_irn_loop(succ);
 712                 if (loop != current_loop) {
 713                         /* a sub-loop? */
 714                         if (get_loop_depth(loop) > get_loop_depth(current_loop)) {
 715                                 find_in_loop(loop, succ);
 716                         } else {
 717                                 /* parent loop: we're not interested in the block */
 718                         }
 719                 } else {
 720                         find_blocks(succ);
 721                 }
 722         }
 723 }
 724
 725 /**
 726  * append an entry to a worklist. WARNING: The entry must not already be in the
 727  * worklist.
 728  */
 729 static void worklist_append(worklist_t *worklist, ir_node *value,
 730                             ir_node *reload_point,
 731                                                         ir_loop *unused_livethrough_loop)
 732 {
 733         worklist_entry_t *entry;
 734
 735 #ifdef EXPENSIVE_CHECKS
 736         foreach_worklist(entry, worklist) {
 737                 assert(entry->value != value);
 738         }
 739 #endif
 740
 741         entry = OALLOCZ(&obst, worklist_entry_t);
 742         entry->value                   = value;
 743         entry->reload_point            = reload_point;
 744         entry->unused_livethrough_loop = unused_livethrough_loop;
 745         list_add_tail(&entry->head, &worklist->live_values);
 746         ++worklist->n_live_values;
 747         assert(worklist->n_live_values <= n_regs);
 748 }
 749
 750 static void push_unused_livethrough(loop_info_t *loop_info, ir_node *value)
 751 {
 752         loop_edge_t *edge;
 753         ++worklist_visited;
 754
 755         /* add the value to all loop exit and entry blocks */
 756         for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
 757                 ir_node            *block
 758                         = get_Block_cfgpred_block(edge->block, edge->pos);
 759                 const block_info_t *info     = get_block_info(block);
 760                 worklist_t         *worklist = info->end_worklist;
 761                 ir_node            *reload_point = NULL;
 762
 763                 if (worklist->visited >= worklist_visited)
 764                         continue;
 765                 worklist->visited = worklist_visited;
 766
 767                 /* TODO: we need a smarter mechanism here, that makes the reloader place
 768                  * reload nodes on all loop exits... */
 769
 770                 worklist_append(worklist, value, reload_point, loop_info->loop);
 771         }
 772         edge = loop_info->entry_edges;
 773         for ( ; edge != NULL; edge = edge->next) {
 774                 ir_node            *entry_block = edge->block;
 775                 const block_info_t *info        = get_block_info(entry_block);
 776                 worklist_t         *worklist    = info->start_worklist;
 777                 ir_node            *pred_block;
 778                 ir_node            *reload_point;
 779
 780                 if (worklist->visited >= worklist_visited)
 781                         continue;
 782                 worklist->visited = worklist_visited;
 783
 784                 pred_block   = get_Block_cfgpred_block(entry_block, edge->pos);
 785                 reload_point = be_get_end_of_block_insertion_point(pred_block);
 786
 787                 worklist_append(worklist, value, reload_point, loop_info->loop);
 788         }
 789
 790         set_irn_link(value, NULL);
 791         ++loop_info->max_register_pressure;
 792 }
 793
 794 static void push_unused_livethroughs(loop_info_t *loop_info)
 795 {
 796         loop_edge_t *edge;
 797
 798         /* we can only push unused livethroughs if register pressure inside the loop
 799          * was low enough */
 800         if (loop_info->max_register_pressure >= n_regs)
 801                 return;
 802
 803         /* find unused livethroughs: register pressure in the loop was low enough
 804          * which means that we had no spills which implies that at every point in
 805          * the loop all*/
 806         for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
 807                 ir_node            *block          = edge->block;
 808                 const block_info_t *info           = get_block_info(block);
 809                 worklist_t         *start_worklist = info->start_worklist;
 810                 ir_node            *exit_block;
 811                 const block_info_t *exit_info;
 812                 worklist_t         *end_worklist;
 813                 struct list_head   *entry;
 814
 815                 if (start_worklist == NULL)
 816                         continue;
 817
 818                 exit_block   = get_Block_cfgpred_block(edge->block, edge->pos);
 819                 exit_info    = get_block_info(exit_block);
 820                 end_worklist = exit_info->end_worklist;
 821
 822                 activate_worklist(end_worklist);
 823                 /* all values contained in the start_worklist, which are not available
 824                  * in the end_worklist, must be unused livethroughs */
 825
 826                 list_for_each(entry, &start_worklist->live_values) {
 827                         worklist_entry_t *wl_entry
 828                                 = list_entry(entry, worklist_entry_t, head);
 829                         ir_node *value = wl_entry->value;
 830
 831                         if (loop_info->max_register_pressure >= n_regs)
 832                                 break;
 833
 834
 835                         if (!worklist_contains(value)) {
 836                                 /* add it to all loop exits */
 837                                 DB((dbg, LEVEL_2, "Found unused livethrough: %+F (regpressure in loop %d)\n", value, loop_info->max_register_pressure));
 838                                 push_unused_livethrough(loop_info, value);
 839                                 /* the value should now be part of end_worklist, so mark it */
 840                                 mark_irn_visited(value);
 841                         }
 842                 }
 843
 844                 deactivate_worklist(end_worklist);
 845         }
 846 }
 847
 848 static void process_block_or_loop(const block_or_loop_t block_or_loop)
 849 {
 850         if (block_or_loop.loop->kind == k_ir_loop) {
 851                 loop_info_t *loop_info = get_loop_info(block_or_loop.loop);
 852
 853                 if (do_push_unused_livethroughs)
 854                         push_unused_livethroughs(loop_info);
 855
 856                 if (loop_info->max_register_pressure > max_register_pressure)
 857                         max_register_pressure = loop_info->max_register_pressure;
 858
 859                 return;
 860         }
 861         process_block(block_or_loop.block, NULL);
 862 }
 863
 864 static void process_loop(ir_loop *loop)
 865 {
 866         int         n_elements = get_loop_n_elements(loop);
 867         int         i, len;
 868         loop_info_t *loop_info;
 869         loop_edge_t *edge;
 870         ir_node     *some_block;
 871
 872         /* first handle all sub-loops */
 873         for (i = 0; i < n_elements; ++i) {
 874                 loop_element element = get_loop_element(loop, i);
 875                 if (*element.kind != k_ir_loop)
 876                         continue;
 877
 878                 process_loop(element.son);
 879         }
 880
 881         /* create a postorder of the blocks */
 882         loop_info = get_loop_info(loop);
 883         edge      = loop_info->entry_edges;
 884         if (edge != NULL) {
 885                 some_block = edge->block;
 886         } else {
 887                 assert(loop == get_irg_loop(current_ir_graph));
 888                 some_block = get_irg_start_block(current_ir_graph);
 889         }
 890
 891         loop_blocks  = NEW_ARR_F(block_or_loop_t, 0);
 892         current_loop = loop;
 893
 894         ir_reserve_resources(current_ir_graph, IR_RESOURCE_BLOCK_VISITED);
 895         inc_irg_block_visited(current_ir_graph);
 896         find_blocks(some_block);
 897         /* for endless loops the end-block might be unreachable */
 898         if (loop == get_irg_loop(current_ir_graph)) {
 899                 find_blocks(get_irg_end_block(current_ir_graph));
 900         }
 901         ir_free_resources(current_ir_graph, IR_RESOURCE_BLOCK_VISITED);
 902
 903         DB((dbg, LEVEL_3, "Block List for loop %p\n", loop));
 904         len = ARR_LEN(loop_blocks);
 905         for (i = 0; i < len; ++i) {
 906                 block_or_loop_t block_or_loop = loop_blocks[i];
 907                 if (block_or_loop.loop->kind == k_ir_loop) {
 908                         DB((dbg, LEVEL_3, " L-%p", block_or_loop.loop));
 909                 } else {
 910                         DB((dbg, LEVEL_3, " B-%+F", block_or_loop.block));
 911                 }
 912         }
 913         DB((dbg, LEVEL_3, "\n"));
 914
 915         max_register_pressure = 0;
 916
 917         /* run1: tentative phase */
 918         tentative_mode = true;
 919         for (i = len-1; i >= 0; --i) {
 920                 process_block_or_loop(loop_blocks[i]);
 921         }
 922
 923         /* run2: tentative phase - should reach fixpoint */
 924         tentative_mode = true;
 925         for (i = len-1; i >= 0; --i) {
 926                 process_block_or_loop(loop_blocks[i]);
 927         }
 928
 929 #ifndef NDEBUG
 930         /* run3: tentative phase - check fixpoint */
 931         tentative_mode               = true;
 932         should_have_reached_fixpoint = true;
 933         for (i = len-1; i >= 0; --i) {
 934                 process_block_or_loop(loop_blocks[i]);
 935         }
 936         should_have_reached_fixpoint = false;
 937 #endif
 938
 939         /* run4: add spills/reloads */
 940         tentative_mode              = false;
 941         do_push_unused_livethroughs = true;
 942         for (i = len-1; i >= 0; --i) {
 943                 process_block_or_loop(loop_blocks[i]);
 944         }
 945         do_push_unused_livethroughs = false;
 946
 947         loop_info->max_register_pressure = max_register_pressure;
 948         DB((dbg, LEVEL_2, "Regpressure in loop %p: %u\n", loop,
 949                                 (unsigned) max_register_pressure));
 950
 951         DEL_ARR_F(loop_blocks);
 952 }
 953
 954 static void fix_block_borders(ir_node *block, void *data)
 955 {
 956         block_info_t *block_info     = get_block_info(block);
 957         worklist_t   *start_worklist = block_info->start_worklist;
 958         int           n_cfgpreds     = get_Block_n_cfgpreds(block);
 959         ir_loop      *loop           = get_irn_loop(block);
 960         int           i;
 961
 962         (void) data;
 963         assert(start_worklist != NULL);
 964
 965         for (i = 0; i < n_cfgpreds; ++i) {
 966                 ir_node      *pred_block      = get_Block_cfgpred_block(block, i);
 967                 block_info_t *pred_block_info = get_block_info(pred_block);
 968                 worklist_t   *end_worklist    = pred_block_info->end_worklist;
 969                 ir_loop      *pred_loop       = get_irn_loop(pred_block);
 970                 bool          is_loop_entry   = false;
 971                 worklist_entry_t *entry;
 972
 973                 assert(end_worklist != NULL);
 974
 975                 if (get_loop_depth(pred_loop) < get_loop_depth(loop)) {
 976                         is_loop_entry = true;
 977                 }
 978
 979                 /* reload missing values */
 980                 activate_worklist(end_worklist);
 981
 982                 foreach_worklist(entry, start_worklist) {
 983                         ir_node *value = entry->value;
 984
 985                         if (!is_loop_entry && entry->unused_livethrough_loop != NULL)
 986                                 continue;
 987
 988                         if (is_Phi(value) && get_nodes_block(value) == block) {
 989                                 value = get_irn_n(value, i);
 990
 991                                 /* we might have unknowns as argument for the phi */
 992                                 if (!arch_irn_consider_in_reg_alloc(cls, value))
 993                                         continue;
 994                         }
 995                         if (worklist_contains(value))
 996                                 continue;
 997                         be_add_reload_on_edge(senv, value, block, i, cls, 1);
 998                 }
 999
1000                 deactivate_worklist(end_worklist);
1001         }
1002 }
1003
1004 /**
1005  * Run the spill algorithm.
1006  *
1007  * @param birg  the graph to run on
1008  * @param ncls  the register class to run on
1009  */
1010 static void be_spill_belady3(be_irg_t *birg, const arch_register_class_t *ncls)
1011 {
1012         ir_graph *irg = be_get_birg_irg(birg);
1013
1014         cls    = ncls;
1015         n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
1016
1017         /* shortcut for register classes with ignore regs only */
1018         if (n_regs == 0)
1019                 return;
1020
1021         worklist_visited = 0;
1022         exec_freq        = be_get_birg_exec_freq(birg);
1023
1024         be_clear_links(irg);
1025         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK
1026                         | IR_RESOURCE_LOOP_LINK);
1027         inc_irg_visited(irg);
1028
1029         obstack_init(&obst);
1030         senv = be_new_spill_env(birg);
1031
1032         assure_cf_loop(irg);
1033         clear_loop_info(get_irg_loop(irg));
1034         irg_block_walk_graph(irg, construct_loop_edges, NULL, NULL);
1035
1036         process_loop(get_irg_loop(irg));
1037
1038         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
1039
1040         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK
1041                         | IR_RESOURCE_LOOP_LINK);
1042
1043         be_insert_spills_reloads(senv);
1044
1045         obstack_free(&obst, NULL);
1046
1047         /* clean up */
1048         be_delete_spill_env(senv);
1049 }
1050
1051 void be_init_spillbelady3(void)
1052 {
1053         static be_spiller_t belady3_spiller = {
1054                 be_spill_belady3
1055         };
1056
1057         be_register_spiller("belady3", &belady3_spiller);
1058         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady3");
1059 }
1060
1061 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady3);